Overview
Local, private, semantic + keyword search over the text messages from an iPhone backup, including what’s in the photos people sent. It’s the note-search recipe — nomic embeddings + BM25, fused with Reciprocal Rank Fusion — pointed at sms.db instead of an Obsidian vault, with a vision model (Florence-2) that captions photos and reads any text in them so they’re searchable too. Nothing leaves the machine: the embedding model, the vision model, and the index all run locally.
How It Works
It reads sms.db and the address book straight from an extracted iPhone backup, then groups messages into conversation windows (single texts like “ok” or “see you at 6” are too short to search on their own). Every photo (and a frame of every video) is captioned with Florence-2, with OCR for screenshots, receipts, and cards, and text pulled from PDFs, contact cards, and Word docs. Search is hybrid: the semantic side finds “lake weekend” when the text actually said “up at the cabin,” and RRF fuses that with keyword hits.
It Also Makes the Book
The same backup can become a keepsake instead of just an index: pass a flag and the tool builds a PDF book of the threads plus a browsable website with the messages and attachment media inline. That export started out as separate standalone scripts — the former Phone SMS to Book Tool — and got folded in here so there’s just one program pointed at the backup.
Current Status
Working and fully local — search, plus the book/website export, all on-machine.